A classification procedure for the automatic
separation of pulse-signals generated by multiple sources
simultaneously active during Partial Discharge (PD)
measurements, is presented in this paper. It is based on the
analysis of signals projected into a 3D space obtained selecting
three different components where the maximum dispersion of the
Normalized Auto-Correlation Function (NACF), is found.
Assuming that the same source can exhibit NACFs having similar
shapes, those which show different shapes are grouped differently
in this space. The DBSCAN algorithm modified to take into
account clusters partially overlapped, is adopted here to separate
the different groups. The relevant phase resolved PD subpatterns
are derived accordingly. Improvements with respect to
the current separation methods are also discussed.
